Issue 4, Volume 3, 2009

151

Statistical simulation of wind speed in Athens, Greece based on Weibull and ARMA models
Kostas Philippopoulos and Despina Deligiorgi
frequencies of the wind speed and subsequently a stochastic ARMA model is trained to generate synthetic time series for each month. The linear ARMA models give comparable results with artificial intelligence statistical methods like the neural networks [28]-[31], and are chosen in this study to assess the ability of statistical models to simulate wind speed data solely based on the autocorrelation of the time series. II. EXPERIMENTAL DATA The study is focused at the metropolitan Athens in Greece and a ten yearly, from January 1993 to December 2002, hourly averaged wind speed time series at the National Observatory of Athens (NOA) is used. The meteorological station of NOA (Fig. 1) is situated in the center of the Athens basin, 9.5km away from the Saronic shoreline, surrounded by mountains in North and East. The wind field in metropolitan Athens is influenced by complex sea-land breeze circulation cells and by katabatic flows from the surrounding mountains.

Abstract Wind Speed simulation and modeling is essential in

the field of wind power estimation and a useful tool in air pollution management and control applications. This study is focused on the simulation of the hourly wind speed sequences of a single station. Under this framework, the Weibull distribution and the autoregressive-moving average ARMA models are employed. The Weibull distribution is fitted to the monthly frequency wind speed distributions. For each month of the year a single ARMA model is trained according to the Box Jenkins methodology. The ARMA models are assessed for their ability to reproduce successfully the main statistical figures of the observed time series. The goodness of fit tests along with the limited percentage error on the observed mean wind speed and standard deviation imply the usefulness of the simulation scheme in generating synthetic wind speed time series for the site under study. ARMA models are found superior in simulating the frequency distributions of wind speed.

KeywordsARMA models, Time series, Wind speed, Weibull

distribution

HE increasingly rising interest in estimating wind power and wind energy potential at a given site, highlights the importance of the statistical simulation of wind speed observations. Early simulations were limited to the distribution fitting of a theoretical probability function like Weibull, Rayleigh or Lognormal to the observed frequency distributions of wind speed [1]-[16]. A detailed review of the probability functions used in wind energy analysis is presented in [17]. Wind speed has a highly auto-correlated nature and the Box-Jenkins methodology [18] is suitable for simulating and forecasting wind speed observations in a specific site [19]-[23] and is proposed and used operationally for simulation and short-term wind speed and power forecasting in wind farms [24]-[27]. In this work, for each month, the Weibull distribution is fitted to the observed
Manuscript received January 26, 2010: Revised version received February 21, 2010. This work was supported by the KAPODISTRIAS research programme of the National and Kapodistrian University of Athens. D. Deligiorgi is with the Physics Department, Division of Environmental Physics and Meteorology, National and Kapodistrian University of Athens 157-84, Athens, Greece (phone: +30 2107276924; fax: +30 210 6018677; email: despo@phys.uoa.gr). K. Philippopoulos is with the Physics Department, Division of Environmental Physics and Meteorology, National and Kapodistrian University of Athens 157-84, Athens, Greece (e-mail: kostasphilippopoulos@yahoo.com).

I. INTRODUCTION

Fig. 1 Area of study and location of NOA meteorological station Wind speed measurements are obtained from a three-cup rotor anemometer, with threshold level 0.2m/s, which is placed at 10m above ground. Wind speed observations are recorded every 10 seconds, averaged initially over ten minutes and then over one hour, to generate the hourly averaged wind speed time series. The main monthly statistical features of the time series (mean wind speed, median, variance, kurtosis and skewness) are presented in Table 1. It is observed that the mean monthly
INTERNATIONAL JOURNAL of ENERGY and ENVIRONMENT

Issue 4, Volume 3, 2009

152

wind speed exhibits two maxima, one in July, which is attributed to the etesian winds and one during the cold period of the year. The mean monthly wind speed frequency distributions have common distribution patterns and they are right-skewed. Table 1 Monthly wind speed descriptive statistics Month Mean Median Variance Kurtosis Skewness Jan 3.318 2.60 5.756 1.477 1.269 Feb 3.537 2.80 6.431 0.854 1.140 Mar 3.781 3.20 6.714 1.662 1.206 Apr 3.026 2.50 3.914 0.681 0.972 May 3.323 2.90 4.906 0.808 0.986 Jun 3.595 3.20 5.261 0.791 0.899 Jul 3.743 3.40 5.862 0.401 0.846 Aug 3.672 3.20 5.564 -0.254 0.723 Sep 3.127 2.50 4.465 1.000 1.117 Oct 3.141. 2.40 5.476 1.347 1.317 Nov 3,338 2.50 6.321 1.518 1.320 Dec 3.401 2.70 6.106 1.540 1.257 III. WEIBULL DISTRIBUTION MODELING Initially the Weibull distribution is fitted to the frequency distributions of the wind speed observations. The Weibull distribution family of curves is a special case of Pearson Type III distributions and is formulated by the probability density function f(u):
f (u ) = k u cc
k 1

2 = c 2 1 +

2 1 2 1 + k k

(3)

Several methods are proposed to determine the shape and scale parameters of the Weibull distribution function [32][34]. In this study the linear regression methodology is adopted and the parameter values of the fitted Weibull distributions are presented in Table 2. The values of k are close to 1, ranging from 1.2380 to 1.4238. This fact implies the high variability of the wind field at metropolitan Athens. The goodness of fit of the Weibull distribution to the frequency distributions of the observed data is assessed by the correlation coefficient (R2) and the Root Mean Square Error (RMSE). The observed and the theoretical values of the mean wind speed and variance are compared by calculating their relative % errors (Table 2). The high correlation coefficient values, ranging from 0.866 to 0.922, along with the low RMSE and % errors, verify that the Weibull distribution fits the data reasonably well. IV. AUTOREGRESSIVE MOVING AVERAGE ARMA(P,Q)
MODELING

u k exp c

(1)

Autoregressive Moving Average ARMA(p,q) models are a group of linear stochastic models which are classified in three categories. The purely autoregressive AR(p) models, the moving average MA(q) models and the mixed ARMA(p,q) models, which are a combination of the autoregressive and moving average processes. At a particular time, the value of the time series in an ARMA(p,q) process is generated by the equation:
x t = + 1 x t 1 + 2 x t 2 + ...... + p x t p + t 1 t 1 2 t 2 ...... q t q

where k is the dimensionless shape factor, c the scale parameter in m/sec and u the wind speed in m/sec. The mean wind speed and the variance 2, in terms of k and c may be calculated using the gamma function , from the following expressions:
1 u = c1 + k

(4)

(2)

where , 1,2, , p, 1,2,, q are the parameters of the mixed ARMA(p,q) model. A three phase methodology framework is proposed by Box and Jenkins [18], which constitutes the identification of a possible ARMA(p,q) model, its parameter estimation phase and its diagnostic checking procedure. % error on 2 10.88% 14.37% 4.66% 9.14% 11.55% 7.08% 8.99% 15.86% 12.20% 13.97% 15.85% 8.59%

Month January February March April May June July August September October November December

k 1.2681 1.3003 1.3228 1.3494 1.3380 1.3965 1.3916 1.4238 1.3488 1.2420 1.2666 1.2380

Table 2: Goodness of fit and Weibull distribution parameters % error on u RMSE c (m/s) R2 3.0719 0.898 0.00063 14.05% 3.2763 0.903 0.00055 14.45% 3.6019 0.922 0.00036 12.32% 2.7452 0.866 0.00657 16.80% 3.0039 0.903 0.00067 16.96% 3.3425 0.914 0.00047 15.22% 3.4786 0.900 0.00053 15.22% 3.3406 0.885 0.00066 17.29% 2.8810 0.882 0.03239 15.49% 2.8731 0.870 0.00095 14.68% 3.1240 0.874 0.00076 13.10% 3.1154 0.901 0.00060 14.49%

INTERNATIONAL JOURNAL of ENERGY and ENVIRONMENT

Issue 4, Volume 3, 2009

153

A. Transformation and Standardization The application of ARMA models requires the modeled data to be stationary and normally distributed. As it is already proved, wind speed is distributed according to the Weibull distribution. Furthermore, wind speed time series are nonstationary, exhibiting seasonal and diurnal variations (Fig. 2).

where WS '(h, m) the expected mean value of hour h for the month m of WS and 2 (h, m ) the variance of hour h for the month m, given by the formulas:
WS '(h, m ) = 1 YM

WS ' (h, d , m, y )
y =1 d =1

(7)
2

2 (h, m ) =

1 YM

[WS ' (h, d , m, y ) WS '(h, m)]

Y M y =1 d =1

(8)

The observed hourly wind speed data have been transformed and standardized to a dimensionless and normally distributed variable. The normality check of WS'* is illustrated in Fig. 3.

Fig. 2 Diurnal patterns of the hourly mean wind speed and standard deviation for March (a) and October (b) Wind speed time series can be transformed to a normally distributed variable by raising each observation to an appropriate index x. The most efficient method determining the index x, is the skewness method [35], which evaluates the symmetry of the distribution, by using the formula:
S k ( x) = 1 Y M D

y =1 d =1

WS x (h, d , m, y ) WS x (m) (5) (m) h =1

D

Fig 3 Normality check for the transformed and standardized wind speed for March (a) and October (b) B. Identification Phase The order of the autoregressive process p and the order q of the moving average process, are estimated during the identification phase. The analysis of the autocorrelation (ACF) and the partial autocorrelation (PACF) functions give an initial estimation of the stochastic process that generated the observed data. For a pure autoregressive model AR(p), the ACF gradually decreases to zero, while the PACF is zero after p lags. For a pure moving average process MA(q), the ACF is zero after q lags and the PACF is decreases geometrically after q lags. For a mixed process ARMA(p,q) the ACF decreases exponentially after lag p and is zero after q lags. For each month, the correlograms (ACF and PACF) of WS'* have the same features, exhibiting a slow exponential decrease (Fig. 4). Furthermore, after the second or third lag, the partial autocorrelation functions are close to zero, implying that the data may be modeled by a low order ARMA(p,q) process.

The method requires iterative calculation and the selected value of index x results to a symmetric distribution (i.e. Sk = 0). By the end of this step, the observed time series WS have been transformed to a normally distributed transformed variable WS'. Seasonal non-stationarity is adequately removed by choosing a monthly scale for the stochastic modeling and standardization is required to remove diurnal non-stationarity [36] The standardization is performed to the WS using the formulas [37]:

WS '* (h, d , m, y ) =

WS ' (h, d , m, y ) WS ' (h, m ) (h, m )

(6)

INTERNATIONAL JOURNAL of ENERGY and ENVIRONMENT

Issue 4, Volume 3, 2009

154

the model coefficients , i and i along with the variance of the residuals e2 can be estimated. Table 3 illustrates the proposed model and the estimated values of the parameters. The stationarity and the invertibility of each model, based on these parameters, are checked. For an ARMA(2,1) model, the following conditions must be fulfilled respectively [18]:

1 + 2 < 1 2 1 < 1 1 < 2 < 1

1 < 1 < 1

(10)

The stationarity and the invertibility of an ARMA(2,2) model are ensured when the estimated parameters satisfy the following conditions:

Fig. 4 Autocorrelation and partial autocorrelation functions for the transformed and standardized variable for March (a&b) and October (c&d) For the selection of the appropriate class of the ARMA models the Bayesian Information Criterion (BIC) is employed. The BIC criterion considers the principle of parsimony, which is essential according to the Box Jenkins methodology, and responds to the following expression:
2 BIC = (DMY ) ln e + T ln (DMY )

1 + 2 < 1 2 1 < 1 1 < 2 < 1

1 + 2 < 1 2 1 < 1 1< 2 <1

(11)

All proposed models were found to be stationary and invertible.

( )

(9)

where D the number of observations in a day, M the number of days in a given month, Y the number of years of the observations, e2 the variance of the residuals and T the total number of the parameters estimated, equal to the order of the ARMA model T=p+q. The BIC criterion is employed to the suggested group of models from the visual analysis of the correlograms. The selected model during the identification phase is the one that minimizes the BIC criterion. An ARMA(2,1) model is proposed for each month except April, October and November where an ARMA(2,2) is selected. .

D. Diagnostic Checking Phase The objective of the diagnostic checking phase is to reveal any lack of fit of the proposed models and diagnose its cause. If the fitted model is adequate then the autocorrelations of the residuals should be uncorrelated and normally distributed and thus their autocorrelations coefficients rk() have to be random and close to zero. Portmanteau Lack-of-Fit test, assesses weather a group of autocorrelations of a time series are random. For this study the first 15 autocorrelations of the residuals are used and the Box-Pierce statistic is calculated:

Q=N

r ( )
2 k k =1

(12)

C. Parameter Estimation Phase Once the provisional values of p and q have been identified, Table 3 ARMA(p,q) model coefficients Month Model 1 2 1 2 e2 January ARMA(2,1) 0.0179 1.34710 -0.38310 0.61965 0.24775 February ARMA(2,1) -0.0002 1.43760 -0.46619 0.70069 0.25835 March ARMA(2,1) 0.0008 1.44046 -0.47070 0.70345 0.27098 April ARMA(2,2) -0.0054 1.72837 -0.73222 1.04146 -0.08454 0.36771 May ARMA(2,1) -0.00150 1.54192 -0.56135 0.81950 0.31995 June ARMA(2,1) -0.0302 1.67530 -0.68082 0.92773 0.31040 July ARMA(2,1) 0.0120 1.63033 -0.63769 0.90658 0.31479 August ARMA(2,1) 0.00310 1.60970 -0.61840 0.90007 0.33662 September ARMA(2,1) -0.00530 1.50437 -0.52133 0.82960 0.34016 October ARMA(2,2) -0.00172 1.69544 -0.70250 1.00313 -0.10057 0.26526 November ARMA(2,2) -0.00413 1.64020 -0.65467 0.96817 -0.12613 0.27094 December ARMA(2,1) 0.00355 1.34329 -0.37881 0.62379 0.25099 The hypothesis of randomness is accepted when the Q-statistic freedom. The results of the Portmanteau Lack-of-Fit test are follows the chi-square distribution, with K (p+q) degrees of presented in Table 4.
INTERNATIONAL JOURNAL of ENERGY and ENVIRONMENT

where N is the number of observations, K the number of the first studied autocorrelations and rk() the residuals autocorrelations.

Issue 4, Volume 3, 2009

155

Table 4 Q statistic values for each month and x2 critical values Month Q statistic x2 critical value January 16,472 27,587 February 19,426 27,587 March 18,265 27,587 April 25,571 26,296 May 26,226 27,587 June 27,394 27,587 July 22,067 27,587 August 16,085 27,587 September 26,099 27,587 October 26,204 26,296 November 26,114 26,296 December 17,090 27,587 The computed Q values for each month were found to be lower than the critical values of the x2 distribution, indicating that the proposed models is accepted at a significance level of 5%. Some researchers [37],[38] follow an alternative method for the diagnosing checking of the models. Their approach is based on the study of the autocorrelations of the residuals for a given number of lags, stating that if the rk2() are uncorrelated then the 95% of the rk2() should be within the

V. SIMULATION RESULTS In order to check the validity of the ARMA models, the synthetic time series are compared with the observed wind speeds for each month of the year. In Fig. 8 the time series are illustrated for 2002, while in Fig. 6, the observed and the synthetic time series are compared with the scatter diagram for the complete 10-year study period. The overall correlation coefficient is high (R = 0.91) while for high wind speeds, both figures illustrate that the models underestimate the observed wind speeds.

2 / N boundaries. The first 20 autocorrelations of the

residuals are calculated for each month and presented in the Fig. 5 along 2 / N error boundaries. Fig 6 Comparison of the observed and simulated wind speed for the complete time series The most important statistical feature of a time series is its autocorrelation. A statistical model reproduces accurately the time series values when the autocorrelation coefficients of the observed and simulated time series are similar. In our case, for both time series the first 8 autocorrelation coefficients were calculated and compared for each month. Fig 7 illustrates the comparison for March and October. For all months, a slight overestimation for the estimated autocorrelation coefficients is observed.

Fig 5 Autocorrelation function of the residuals for March (a) and October (b) Both diagnostic checking techniques lead to the same findings, proving that the proposed models are adequate for the simulation of the wind speed.

Fig 7 Comparison of the observed and simulated autocorrelation functions for the first 8 lags for March (a) and October (b)

INTERNATIONAL JOURNAL of ENERGY and ENVIRONMENT

Issue 4, Volume 3, 2009

156

Fig 8 Comparison of the observed and simulated time series for 2002 Furthermore, the comparison of the main statistical characteristics (Table 5), proves the ability of the ARMA models to generate and reproduce accurately the wind speed at the National Observatory of Athens. The comparison is based on the mean monthly wind speed, its standard deviation and on the two first autocorrelation coefficients. The percentage errors for each of the above statistical figures are calculated from the following expression:
error (%) = (Obs Sim ) 100 Obs

(13)

Limited errors are observed for all statistical figures except for the standard deviation. Furthermore, the lower mean monthly wind speed percentage errors are observed for the summer months and the higher during autumn. The comparison of the frequency distributions of the observed and the synthetic time series (Fig. 9) prove that the fit of the ARMA models to actual data is very promising. This is important especially for the lower wind speeds, where the Weibull distribution is incapable of simulating the observed distribution frequencies. Fig 9 Observed and simulated frequency distributions for March (a) and October (b)

INTERNATIONAL JOURNAL of ENERGY and ENVIRONMENT

Issue 4, Volume 3, 2009

157

Table 5 Comparison of the observed and simulated main statistical figures along with the % errors on the monthly mean, standard deviation and the first and second autocorrelation coefficients. u u error Month r1 r2 r1error r2error 2 2error Obs 3.318 5.756 0.916 0.844 January 3.986 25.010 2.620 4.265 Sim 3.186 4.316 0.940 0.880 Obs 3.537 6.431 0.915 0.833 February 4.042 25.042 2.186 3.721 Sim 3.394 4.820 0.935 0.864 Obs 3.781 6.714 0.900 0.815 March 3.682 26.022 2.889 4.417 Sim 3.642 4.967 0.926 0.851 Obs 3.026 3.914 0.872 0.746 April 3.724 28.081 4.128 6.032 Sim 2.913 2.815 0.908 0.791 Obs 3.323 4.906 0.893 0.785 May 3.049 23.882 3.024 4.968 Sim 3.221 3.734 0.920 0.824 Obs 3.595 5.261 0.897 0.782 June 2.535 21.416 2.341 4.348 Sim 3.504 4.134 0.918 0.816 Obs 3.743 5.862 0.899 0.786 July 2.563 20.883 2.781 5.089 Sim 3.647 4.638 0.924 0.826 Obs 3.672 5.564 0.892 0.777 August 2.844 21.850 3.139 5.534 Sim 3.567 4.348 0.920 0.820 Obs 3.127 4.465 0.900 0.798 September 3.863 26.150 3.333 5.764 Sim 3.006 3.297 0.930 0.844 Obs 3.141 5.476 0.920 0.845 October 4.580 24.582 2.826 4.142 Sim 2.997 4.130 0.946 0.880 Obs 3.338 6.321 0.918 0.846 November 4.753 27.209 3.050 4.374 Sim 3.180 4.601 0.946 0.883 Obs 3.401 6.106 0.917 0.843 December 4.234 25.617 2.617 4.389 Sim 3.257 4.542 0.941 0.880 their relative errors. Furthermore the synthetic time series are proved to reproduce accurately the autocorrelation VI. CONCLUSION In this study a theoretical distribution function and a dependence, which is the most important feature in wind stochastic modeling approach are employed for the simulation speed time series. A slight overestimation of no statistical of the hourly wind speed observations at a single station in significance is observed for the first eight autocorrelation metropolitan Athens. Weibull distribution, which is the most coefficients. The fit of the ARMA model to the frequency distributions frequently used distribution in wind speed statistical analysis, of the observed data is superior compared to Weibull models, is found to model satisfactory the observed relative frequency distributions of wind speed for each month. Although high especially for low wind speeds. ARMA models are found correlation coefficient values and low RMSE errors are found capable in generating synthetic time series with identical for each month, the Weibull distribution does not model statistical characteristics with the measured data. These stochastic models may be used as a weather wind effectively the low wind speeds and does not take into account speed generators of one month sequences that represent the the autocorrelation feature of wind. The analysis of the statistical characteristics of the time actual statistical characteristics of the 10 years time series of series, based on the autocorrelation and the partial wind speed data for each month. The simulation can be autocorrelation functions, imply that the simulation should be especially useful in generating missing wind speed data for based on a low order ARMA process. The Box Jenkins the National Observatory of Athens and in air pollution methodology is employed and the time series are transformed modeling and control. Furthermore, such a simulation is and standardized, generating a dimensionless and normal important in energy conversion studies in wind energy distributed variable. A different ARMA model is trained for applications. each month and an ARMA(2,1) is proposed for each month except April, October and November where an ARMA(2,2) is selected. The comparison of the observed and simulated time series proves the ability of the model to generate the wind speed at the NOA station. The synthetic time series follow closely the observed wind speed. The comparison is based on the monthly mean wind speed and standard deviation and REFERENCES [1] [2]
[3] A. Garcia, J. L. Torres, E. Prieto, A. De Francisco, Fitting wind speed distributions : a case study, Solar Energy, vol. 62, no 2, pp. 139 144, 1998. M. Y. Sulaiman, A. M. Akak, M. A. Wahab, A. Zakaria, Z. A. Sulaiman, J. Suradi, Wind characteristics of Oman, Energy, vol 27, pp. 35-46, 2002. A. N. Celik, A statistical analysis of wind power density based on the Weibull and Rayleigh models at the southern region of Turkey, Renewable Energy, vol 29, no4, pp. 593 604, Apr. 2004.

INTERNATIONAL JOURNAL of ENERGY and ENVIRONMENT

Issue 4, Volume 3, 2009

[4] [5] [6] [7] [8] [9] [10] [11] [12] [13] E. S. Takle, J. M. Brown, Note on the use of Weibull statistics to characterize wind-speed data, J. Appl. Meteor., vol. 17, no. 4, pp. 556 559, Apr. 1978. S. Binova, R. Burlon, C. Leone, Hourly wind speed analysis in Sicily, Renewable Energy, vol. 28, pp. 1371 1385, 2003. J. L. Torres, A. Garcia, E. Prieto, A. De Franscisco, Characterization of wind speed data according to wind direction, Solar Energy, vol 66, no1, pp. 57 64, 1999. S. Mather, K. P. Pandey, A. V. Kumar, Analysis of wind regimes for energy estimation, Renewable Energy, vol. 25, pp. 381 399, 2002. K. Rosen, R. Van Buskirk, K. Garbesi, Wind energy potential of coastal Eritrea: An analysis of sparse wind data, Solar Energy, vol. 66, no3, pp.201 213, 1999. I. Y. F. Lun, J. C. Lam, A study of Weibull parameters using long term wind observations Renewable Energy, vol. 20, pp. 145 153, 2000. B. Ozerdem, M. Turkeli, An investigation of wind characteristics on the campus of Izmir institute of technology, Turkey, Renewable Energy, vol. 28, pp. 1013 1027. J. Z. Yim, J G Lin, C H Hwang, Statistical properties of the wind field at Taichung harbor, Taiwan, J. Wind Eng. And Ind. Aerodyn., vol. 83, pp. 49 60, 1999. C. P. Jacovides, C. Theophilou, F. S. Tymvios, S. Pashiardes, Wind statistics for coastal stations in Cyprus, Theor. Appl. Climatol., vol 72, pp. 259 263, 2002. A. Shamilov, I. Usta, Y. M. Kantar, The Distribution of Minimizing Maximum Entropy: Alternative to Weibull distribution for Wind Speed, WSEAS Trans. on Mathematics, vol. 5, issue 6, pp. 695 700, June 2006. A. Zaharim, S. K. Najid, A. M. Razali, K. Sopian, The suitability of statistical distribution in fitting wind speed data, WSEAS Trans. on Mathematics, vol. 7, issue 12, pp. 718 727, Dec. 2008. I. Fyrippis, P. J. Axaopoulos, G. Panayiotou, Analysis of wind potential and energy production in Naxos island, Greece, WSEAS Trans. on Power Sys., vol. 3, issue 8, pp. 567 576, Aug. 2008. S. K. Najid, A. Zaharim, A. M. Razali, M. S. Zainol, K. Ibrahim, K. Sopian, Analyzing the east coast Malaysia wind speed data, Int. J. of Energy and Env., vol. 3, issue 2, pp. 53 60, 2009. J. A. Carta, P. Ramirez, S. Velazquez, A review of wind speed probability distributions used in wind energy analysis case studies in the Canary islands, Renewable and Sustainable energy Reviews, vol. 13, pp. 933 955, 2009. G. Box, G. Jenkins, G. Reinsel, Time Series Analysis, Forecasting and Control, Upper Saddle River NJ: Pentice Hall, 1994, pp. 181 326. M. Blanchard, G. Desrochers, Generation of autocorrelated wind speeds for wind energy conversion system studies, Solar Energy, vol. 22, no. 6, pp. 571 579, 1984. B. Brown, R. W. Katz, A. H. Murphy, Time series models to simulate and forecast wind speed and wind power, J. Clim. Appl. Meteor., vol. 23, pp. 1184 1195, Aug. 1984. J. L. Torres, A. Garcia, A. De Francisco, Forecast of hourly average wind speed with ARMA models in Navarre (Spain), Solar Energy, vol. 79, issue 1, pp. 65 77, July 2005. L. Kamal, Y. Z. Jafri, Time series models to simulate and forecast hourly averaged wind speed in Quetta, Pakistan, Solar Energy, vol. 61, no. 1, pp. 23 32, 1997. A. R. Daniel, A. A. Chen, Stochastic simulation and forecasting of hourly average wind speed sequences in Jamaica, Solar Energy, vol. 46, no. 1, pp. 1 11, 1991. M. Milligan, M. Schwartz, Y. H. Wan, Statistical wind power forecasting for U.S. wind farms, presented in Windpower 2003, Austin, 2003. Y. K. Wu, J.S. Hong, A literature review of wind forecasting technology in the world, in 2007 Power Tech, 2007, Laussane, pp. 504 509. T.H.M. Nielsen, Experiences with statistical methods for wind power prediction, in Proc. Eur. Wind Ener. Conf., Nice, 1999, pp. 1066 1069. C.L.L. Ran, Short-term wind speed forecasting model for wind farm based on wavelet decomposition, in Proc. 3rd Int. Conf. in Electric Utility Deregulation and Reconstructing and Power Technologies, 2008, Nanjuing, pp. 2525 -2529.

158
[28] M.A. Mohamed, S. Rehman, T.O. Halawani, A neural network approach for wind speed prediction, Renewable Energy, vol. 13, no. 3, pp. 345 354, 1998. [29] A. More, M.C. Deo, Forecasting wind with neural networks, Marine Structures, vol. 16, pp. 35 49, 2003. [30] M. L. L. Shiyan, J. Chuanwen, L. Hongling, Z. Yan, A review on the forecasting of wind speed and generated power, Renewable & Sustainable energy Reviews, vol. 13, pp. 915 920, 2009. [31] A. Sfetsos, A comparison of various forecasting techniques applied to mean hourly wind speed time series, Renewable Energy, vol. 21, pp. 23 35, 2000. [32] J. V. Seguro, T.W. Lambert, Modern estimation of the parameters of the Weibull wind speed distribution for wind energy analysis, Journal of Wind Engineering and Industrial Aerodynamics, vol. 85, pp. 75 84, 200. [33] S. A. Akdag, A. Dinler, A new method to estimate Weibull parameters for wind energy applications, Energy Conversion and Management, vol. 50, pp. 1761 1766, 2009. [34] K. Conradsen, L. B. Nielsen, L. p. Prahm, Review of Weibull statistics for estimation of wind speed distributions, J. Clim. Appl. Meteor., vol. 23, pp. 1173 1183, Aug. 1984. [35] A. Balouktsis, D. Tsanakas, G. Vachtsevanos, Stochastic simulation of hourly and daily average wind speed sequences, Wind Engng, vol. 6, pp. 219 228, 1986. [36] H. Nfaoui, J. Buret, A. A. M. Saygh, Stochastic simulation of hourly average wind speed sequences in Tangiers, Solar Energy, vol. 53, no. 3, pp. 301 314, 1996. [37] P. Poggi, M. Muselli, G. Notton, C. Cristofari, A. Louche, Forecasting and simulating wind speed in Corsica by using an autoregressive model, Energy Conversion and Management, vol. 44, pp. 3177 3196, 2003. [38] Chatfield C., The Analysis of Time Series an Introduction, London U.K.: Chapman & Hall, 1989, pp. 49 65.

[14] [15] [16] [17]

[18] [19] [20] [21] [22] [23] [24] [25] [26] [27]

INTERNATIONAL JOURNAL of ENERGY and ENVIRONMENT